Harness Engineering for AI Agents
Fikayo Adepoju
Trained 12000+ learners on AI Agents
Everyone's shipping agents. Almost no one is engineering the harness.
The model is the easy part. The hard part is the harness: the workspace you give your agent, the code execution it runs in, the memory and search it reaches for, the way you fight context rot, and the observability that tells you what actually went wrong. This is the layer that decides whether your agent works in production or quietly fails in front of users.
Harness engineering is a discipline. It has patterns, anti-patterns, and load-bearing decisions that most teams make accidentally on a Tuesday afternoon and then live with for a year.
This one-day workshop is a deep, framework-agnostic walkthrough of that discipline. One harness layer per hour, each with a concept walkthrough and demo, drawn from real harnesses running in production today.
Workshop agenda
- 10:30AM EDT
Foundations: Model + Harness (30 min)
Understand the equation Agent = Model + Harness, map the six harness layers, and learn the working-backwards design method that drives every architectural decision in the rest of the day.
- 11:00AM EDT
Layer 1: The Filesystem (45 min)
Learn why the filesystem is the foundational harness primitive and how to wire workspace tools, git versioning, and the AGENTS.md pattern to give your agent durable cross-session memory.
- 11:45AM EDT
15 Min Break
Pause and recharge
- 12:00PM EDT
Layer 2: Code Execution (45 min)
Move past fixed tool sets by giving your agent bash as a meta-tool. Master the full ReAct loop with the safety guardrails (allow-lists, timeouts, stderr capture) every production system needs.
- 12:45PM EDT
30 Min Meal Break
Pause and recharge
- 1:15PM EDT
Layer 3: Sandboxing (45 min)
Replace unsafe local execution with containerised sandboxes. Learn to pre-configure environments in Docker, wire self-verification loops, and isolate the network surface your agent runs on.
- 2:00PM EDT
Layer 4: Memory & Search (45 min)
Solve the knowledge gap problem with a layered memory strategy: durable AGENTS.md memory, live web search, MCP for structured sources, and vector retrieval for longer projects.
- 2:45PM EDT
15 Min Break
Pause and recharge
- 3:00PM EDT
Layer 5: Context Management & Long-Horizon Execution (45 min)
Fight context rot with compaction, tool-call offloading, and progressive tool disclosure. Then enable long-horizon work with planning, the Ralph Loop, and parallel subagent spawning.
- 3:45PM EDT
Layer 6: Observability & Harness Optimization (45 min)
Make agent failures findable. Learn to instrument the harness with full traces, try a lightweight eval suite, and run the benchmark-diagnose-fix loop that turns observability into improvement.
Learn directly from Fikayo
Fikayo Adepoju
I've taught thousands of engineers the patterns that hold up in production
Who this workshop is for
AI/ML engineers already shipping LLM features who want to move from "it works" to "it works in production"
Backend and software engineers moving into agent work who want the mental models before they pick a framework
Tech leads and staff engineers designing agent systems who need a defensible architecture and a vocabulary to evaluate vendor claims
What's included
Live sessions
Learn directly from Fikayo Adepoju in a real-time, interactive format.
Lifetime access
Go back to course content and recordings whenever you need to.
Community of peers
Stay accountable and share insights with like-minded professionals.
Certificate of completion
Share your new skills with your employer or on LinkedIn.
Maven Guarantee
Your purchase is backed by the Maven Guarantee.
Frequently asked questions
$500
USD
2 cohorts